University of Wollongong
Research Online
Faculty of Commerce - Papers (Archive)
Faculty of Business
2011
The impact of job loss on family dissolution
Denise Doiron
University of New South Wales
Silvia Mendolia
University of Wollongong, [email protected]
Publication Details
Doiron, D. & Mendolia, S. (2011). The impact of job loss on family dissolution. Journal of Population Economics, 25 (1), 367-398.
Research Online is the open access institutional repository for the University of Wollongong. For further information contact the UOW Library:
[email protected]
The impact of job loss on family dissolution
Abstract
The impact of involuntary job displacements on the probability of divorce is analysed using discrete duration
models. The analysis uses the sample of couples from the British Household Panel Survey and distinguishes
between types of displacements. Results show that couples in which the husband experiences a job loss are
more likely to divorce. Redundancies have small, positive, often insignificant and short-lived effects while
dismissals and temporary job endings have larger positive impacts. This is consistent with the interpretation of
redundancies as capturing negative income shocks while other types of job loss also convey new information
about potential future earnings and match quality.
Keywords
era2015, loss, family, impact, dissolution, job
Disciplines
Business | Social and Behavioral Sciences
Publication Details
Doiron, D. & Mendolia, S. (2011). The impact of job loss on family dissolution. Journal of Population
Economics, 25 (1), 367-398.
This journal article is available at Research Online: http://ro.uow.edu.au/commpapers/1038
The impact of job loss on family dissolution 1
by
Denise Doiron and Silvia Mendolia
School of Economics,
Australian School of Business Building
University of New South Wales,
Sydney, NSW 2052 Australia
Abstract
The impact of involuntary job displacements on the probability of divorce is analysed using
discrete duration models. The analysis uses the sample of couples from the British Household Panel
Survey and distinguishes between types of displacements. Results show that couples in which the
husband experiences a job loss are more likely to divorce. Redundancies have small, positive, often
insignificant and short-lived effects while dismissals and temporary job endings have larger
positive impacts. This is consistent with the interpretation of redundancies as capturing negative
income shocks while other types of job loss also convey new information about potential future
earnings and match quality.
JEL Codes: J12, J60, J63
Keywords: job loss, divorce, marriage duration
1
Corresponding author: Silvia Mendolia – University of Aberdeen Business School. Edward Wright Building.
Dunbar Street. Old Aberdeen. AB24 3QY
Email address: [email protected]
1
1. Introduction1
The aim of this paper is to analyse the relationship between labour market outcomes and family
well-being. Specifically, we focus on the impact of job losses on family dissolution. The
involuntary termination of employment usually leads to lower earnings and the stress created by the
negative income shock can increase the probability of family dissolution. Moreover, a job
displacement may signal individual traits that impact negatively on future earnings or on the quality
of the match more generally. Again this suggests a positive relationship between family dissolution
and job displacement. While policies aimed at reducing the earnings’ shock from job losses may
alleviate stress on the couple in the case of the former effect, they will have less impact if the latter
effect is the dominant one.
In recent decades, family and marriage characteristics have dramatically changed; divorce rates
have risen and marriage rates fallen. Fertility has declined, longevity increased and cohabitation has
emerged as an important institution, often as a substitute for marriage. Economic models that
explain why people marry and remain together have been developed in an attempt to explain these
changes (Ermisch, 2003). The economic approach to marriage is based on the assumption that
couples marry and stay married when the net gains from marriage are greater than those from
remaining single. According to the traditional models of household economics, starting from Gary
Becker’s Treatise on Family, these gains result mostly from gender specialization (especially when
raising children) and sector-specific investments in human capital (Becker, 1974 and Becker 1991).
Fundamental changes in the technology of household production and of birth control and increased
female labour market participation have altered the returns to household specialization and,
according to the traditional model, reduced the opportunity cost of marriage. More general models
of household production have shifted attention away from specialization and the division of labour,
to the benefits of joint consumption and a positive match between husbands’ and wives’
preferences (Lundberg, 2005 and Lam, 1988).
The analysis of the relationship between job loss and family dissolution is particularly appropriate
in this context. Given the increased complexity of the marriage relationship, individuals are likely
to spend more time searching for a good match on the marriage market (Gould and Paserman,
2003). We also expect partners to re-evaluate the benefits from an existing marriage more
frequently. A husband’s involuntary job loss can lead to such a re-examination of the partnership as
2
it can affect both the contemporaneous value of the marriage (compared to its alternatives) and
signal the likelihood of undesirable traits and lower future earnings.
To fix ideas, consider a stylised model of marriage and divorce involving an initial match quality
that is known at the time of marriage and the evolution of this match quality over time. Partners
stay together if the match quality (including future expected benefits of the partnership) is high
relative to outside options. The dynamic process underlying match quality will depend on various
factors including choices made by the partners (e.g. having children) and state dependence whereby
the match quality is causally affected by the duration of the match to date. Match quality will also
depend on the occurrence of various shocks some of them having only contemporaneous effects
while others generate more persistent impacts. A job loss can cause a contemporaneous shock in
the loss of earnings and an increased stress on the union. It can also cause a more persistent effect if
the earnings loss is long lasting or if the future expected match quality is revised. The latter will
occur when the job loss is seen as a signal of lower future match benefits. This framework underlies
the empirical approach adopted in the paper.
It is important to understand the reason for the termination of the employment spell in order to
evaluate what information this event may convey regarding the partner’s suitability. An involuntary
and exogenous displacement causes an income shock, but does not convey new information about
the partner’s characteristics. On the other hand, a “person-specific” dismissal is likely to be caused
at least in part by the individual’s characteristics and behaviour. Papers studying the effects of
layoffs on future earnings and probabilities of employment support these hypotheses. Job losses
from plant closures (Gibbons and Katz, 1991; Doiron, 1995) or redundancies (Arulampalam, 2001)
have a smaller effect on future earnings than other types of involuntary displacements.
In this paper, data from the British Household Panel Survey (BHPS) are used to analyse the effects
of involuntary job losses experienced by the husband on the probability of marital dissolution. We
take into account the length of the union and we estimate discrete time duration models. Individual
and family characteristics are included and in some specifications, unobserved, time-invariant and
match-specific random effects are modelled. The reasons for the termination of the employment
spell are used to distinguish between different types of job losses: dismissals, redundancies and
temporary job endings. While dismissals are more likely to be related to individual traits,
redundancies are based on the employer’s characteristics and environment and are expected to
represent effects of earnings shocks only.
3
According to the legal definition of redundancy in the UK, this type of job loss should be
uninformative about the individual’s traits. The British legislation is quite explicit that the term
redundancy should not refer to a dismissal caused by an individual worker’s behaviour. (We
discuss the definition of redundancies more extensively in the data section.) Also, the distinction
between types of displacements is supported by a recent study of the BHPS. Arulampalam (2001)
finds that redundancies overall have less of a scarring effect; specifically, she finds that the
earnings loss due to redundancies is about one half of that due to other displacements and 81% of
men made redundant found jobs without any spell of non-employment. In another study of the
BHPS, Borland et al (2000) compare the earnings loss of workers based on the reasons for the
termination of the employment spell. They argue that the institutional system often blurs the
distinction between the different categories and separate displaced workers from industries with
decreasing employment in order to further reduce the potential bias from endogenous variations in
job losses 2.
A job loss that contains a signal of the individual’s characteristics is more likely to be correlated
with unobservable components of match quality and lead to endogeneity bias. Furthermore, this
bias is expected to be positive in the sense that it inflates effects of the job loss variables. Much of
the empirical analysis conducted in the paper is designed to provide evidence on whether the
different types of job loss impact differentially on the probability of divorce. The results all point to
the treatment of redundancies as exogenous events in this context.
There are very few papers looking at the effects of job losses on marital dissolution and to our
knowledge only two of them take into account the reason for displacement. Furthermore these two
papers find conflicting results. Charles and Stephens (2004) analyse US data and find that unlike
layoffs, displacements due to plant closures have no significant effect on divorce 3. On the other
hand, Eliason (2004) finds that factory closures in Sweden do lead to an increase in dissolutions.
Our paper provides new evidence on this topic based on new data and models.
Our empirical framework is closer to that of Charles and Stephens (2004) in the sense that we
estimate duration models that account for marital longevity4. However there are also substantial
differences between the approaches in the two papers. Our models are based on a proportional
hazards framework and the resulting estimates are easier to interpret. We also estimate a variety of
specifications of the hazard and look at selection into the stock and flow samples. Finally,
cohabitations are included in the analysis.
4
Whether the displacement contains a signal of unfavourable individual traits or simply represents
an earnings shock has several implications in the empirical analysis. For example, “personspecific” dismissals are expected to have more severe and longer lasting impacts on divorce
probabilities. This is in fact what we find. A redundancy experienced by the husband has positive
but insignificant effects on the probability of family dissolution. The impact of dismissals is much
higher - an increase of around 154% one year later - and statistically significant. (The average
divorce rate for the sample is between 1 and 2% per year.) Effects of temporary job endings are
generally located between the two. The addition of lags in the job loss variables does not change
these overall results. Random effects specifications also yield similar results, an indication that our
model of the baseline hazard is sufficiently flexible to capture the correlation caused by matchspecific unobservables across time.
Using information on the workforce growth rate by industry, we separate job losses occurring in
declining and growing industries. Displacements in declining industries are less likely to represent
signals of unfavourable individual traits. Results differ considerably by type of job loss;
redundancies from declining industries have a significant positive effect on the probability of
divorce (a 100% increase one year later) while dismissals have a significant effect only in
expanding industries (an increase of 161%). Again results are robust to semi parametric random
effects specifications. These findings are consistent with the view of redundancies as capturing
earnings shocks while dismissals contain signals of a bad match. Earnings shocks are more serious
in declining industries given the difficulty of finding new and equivalent employment while
unfavourable individual traits are more likely explanations of displacements in tight labour
markets.
Turning to other results, there is strong evidence of duration dependence in marital stability. In
general, the longer people have been married, the smaller the probability of family dissolution. This
result is reversed for the intermediate durations (10 to 20 years of marriage) where the probability
of divorce is increasing with time ceteris paribus. The wife’s nonlabour income and age increase
the probability of divorce as do large differences in the partner’s ages. Other regressors are
generally individually insignificant. This is partly due to specification choices; we concentrate on
the effects of job losses and adopt flexible specifications for all other variables.
The paper is organized as follows. The following Section provides an overview of the existing
literature. Section 3 includes a description of the data construction and descriptive statistics.
5
Section 4 discusses the econometric model and Section 5 presents the empirical results. Finally
Section 6 contains concluding comments.
2. Overview of existing literature
We begin the section with a brief discussion of economic models of marriage and divorce.
Although we focus on economic models, it is clear that economic considerations form but part of
the picture and as stated by Weiss and Willis (1997): “A successful theory which is capable of
explaining the data on marriage and divorce must incorporate ideas from sociology, biology and
other fields”. Nonetheless, economic factors have been shown to play a significant role in the
decisions to form and dissolve households.
Becker’s seminal work (1974) forms the first economic framework of marriage and divorce. Two
individuals marry when there is a positive surplus from their union relative to the two remaining
single. As long as they are married, the two individuals maximise a joint expected utility function,
whose arguments are the income or labour earnings received by each spouse (see Weiss, 2000 and
Charles and Stephens, 2004 for more details). The couple divorces when the joint expected utility
of being married is less than the sum of the individual expected utilities from divorce. The expected
utility of divorce includes the probability of remarriage as well as the costs of divorce and the
expected utility of remaining married includes the future option of divorce.
Two general causes for marital instability and divorce are present in this model. First, although the
search for a partner is costly, meetings do occur on a random basis. As a consequence, a union may
become unacceptable if one of the two partners meets a person who would be superior to the
current match. Second, people enter a marriage based on expectations about the match-quality
which depends on the traits of the other spouse. These characteristics may change over time
unexpectedly and cause the spouses to reconsider their initial decision (see Weiss and Willis, 1997
and Boheim and Ermish, 2001). Thus “surprises”, such as unexpectedly high or low income, may
affect marriage dissolution. According to Becker, Landes and Michael (1977), “the majority of
divorces results from uncertainty and unfavourable outcomes”.
A job loss may be considered as an economic “surprise” impacting negatively on the partner’s
future expected earnings. It could also be a signal of characteristics (heretofore unknown) of the
partner that affect his/her suitability as a mate such as reliability or sense of responsibility. Eliason
(2004) underlines that the traits needed to keep a job are partly the same as the traits that make a
6
partner desirable. Hence a job loss may reveal new information about the match quality and lead to
marital dissolution.
An alternative theory of divorce is the family stress theory first elaborated by Hill (1949) and later
by McCubbin and Patterson (1982) 5. A job displacement can be considered as a stressor event and
have an impact on the family’s coping resources potentially leading to a crisis or a resolution. For
example job loss is found to be correlated with alcohol abuse (Catalano et al, 1993) and domestic
violence (Kyriacou at al, 1999). Game theoretic models of family bargaining offer alternatives to
unitary models. In the “divorce threat models” bargaining power depends on the expected utility
outside of marriage (Manser and Brown, 1980 and McElroy and Horney, 1981). In Lundberg and
Pollak (1996), both partners behave noncooperatively and treat divorce as an outside option.
Finally, in a recent study, Matouschek and Rasul (2004) develop stylised models of marriage as an
exclusive contract. In a repeated games context, marriage can act as a commitment device that
fosters cooperation.
In all these models, job displacement plays a natural role in explaining marriage dissolution.
Furthermore, several channels of transmission are expected. A job loss can be a stressor event, a
signal of altered future earnings or more generally future match quality, an indication of shifts in
household bargaining powers, the values of the outside option, and the degree of commitment to
the marriage.
We now turn to empirical studies analysing the effects of job losses. There is a well-established
body of work showing the effects of job displacement on re-employment probabilities and future
earnings. Displaced workers tend to experience reduced employment possibilities, increased job
instability, as well as lower earnings’ profiles (Ruhm, 1991; Jacobsen, Lalonde and Sullivan, 1993;
and Chan and Stevens, 2001). A growing number of studies consider the effects of job loss on the
behaviour of other members of the family. For example Stephens (2001) analyses family
consumption changes after the husband’s job loss; also Ercolani and Jenkins (1999) and Stephens
(2004) focus on wives’ labour supply changes in response to the husband’s job loss 6.
Changes in family labour supply and consumption form only part of the impact of job loss and the
reduction in earnings. Recent work shows substantial impacts of unemployment on mental and
physical health and well-being generally. There is a large empirical psychological literature 7
investigating the impact of unemployment on the incidence of low life satisfaction, depression, low
self-esteem, unhappiness, and even suicide. The negative income shock is but one source of these
7
effects as employment is also a provider of social relationships, one’s identity in society and
individual self esteem (Winkelmann and Winkelmann, 1998). For example, a British study by
Clark and Oswald (1994) shows that unemployed people have much lower levels of mental well
being than those in work and Sullivan and von Watcher (2006) find a significant relationship
between job loss and mortality8.
The effect of job displacements on decisions regarding fertility and marriage forms yet another
dimension of the impact of job loss. These non-pecuniary adjustments cannot be regarded as being
of secondary importance; divorce is ranked as the second most stressful of life events following
death of a family member (Miller and Rahe, 1997). Nevertheless, there is but limited research on
this aspect of the costs of job displacements; furthermore, to date these papers provide inconsistent
evidence.
Jensen and Smith (1990) analyse Danish panel data and find significant effects of job losses on
divorce but only for contemporaneous spells of unemployment. Job losses occurring one or two
years earlier have no impact. These findings raise concerns that reverse causality may be driving
findings of significant effects of job losses when the timing of events is not accounted for.
Information regarding the length of the union is not used in this study.
Weiss and Willis (1997) use US data from the National Longitudinal Study of the High School
Class of 1972 to study the effects of earnings shocks on the probability of divorce. Shocks or
“surprises” are defined as the difference between realized and predicted earnings estimated from
earnings regressions. They show that a positive surprise to the husband’s earnings lowers the
probability of marriage dissolution, while a positive shock in the wife’s earnings raises the chance
of divorce. These results are robust to the inclusion of several controls for match quality. More
recent studies use direct measures of earnings’ shocks. For example, based on the German Socioeconomic panel data, Kraft (2001) analyses the impact of unemployment on married couples’
decision to separate. The husband’s unemployment is found to increase the risk of separation in the
following year and this impact increases with the duration of unemployment.
Using the Panel Study of Income Dynamics, Charles and Stephens (2004) find an increase in the
probability of divorce following a spouse’s job displacement in the first three years. In the last part
of this paper, they compare different job losses and find a significant increase only for layoffs and
not for plant closures. As Charles and Stephens (2004) state “This suggests that information
conveyed about a partner’s non-economic suitability as a mate due to a job loss may be more
8
important than the financial losses in precipitating a divorce.” In contrast, Eliason (2004) finds a
significant negative impact on the marriage’s stability in the long term (up to 13 years after the
displacement) caused by the husband’s or the wife’s job displacement due to a factory closure in
Sweden.
Lastly, in an independent study to ours, Blekesaune (2008) finds a significant increase in the
probability of family dissolution after any form of unemployment (experienced by husbands or
wives). The paper is based on the BHPS and panel data techniques (random effects models) are
used to control for unobserved heterogeneity. One major difference with our analysis is that
Blekesaune does not distinguish between different causes of unemployment.
3. Data construction and descriptive statistics
The analysis uses data collected in the first 14 waves of the British Household Panel Survey
(BHPS), which is a nationally representative sample recruited in September 1991. The survey
contained approximately 10,000 persons (5,500 households) when it was constituted 9. The BHPS is
an indefinite life panel survey and the longitudinal sample consists of members of original
households and their natural descendants. If the original members split off from their household to
form a new family, all the adult members (older than 16) of the new household are included in the
survey and interviewed.
In order to analyse the possible impact of job loss on family dissolution, we firstly construct a
sample of all married or cohabitating couples in the BHPS. A dataset containing consolidated
marital, cohabitation and fertility histories for the 29,065 adults interviewed at least once during the
survey is available together with the original data (see Pronzato, 2007). This dataset provides the
starting and end date of each union. If the union is a marriage, one or both partners can die, they
can get divorced, separated or stay together. If the union is cohabitation, the partners can split, get
married or they can continue cohabitating. In this analysis, we do not distinguish between
marriages and cohabitations 10. If the two partners cohabitate before marriage, we consider the
cohabitation starting date as the union starting date. If there is a separation before the divorce, the
date of separation is considered as the union end date 11.
A divorce binary variable is defined to equal 1 when the end date from the family data set indicates
a separation, a divorce or a split (for cohabitating partners) and when this is the last time the couple
is observed being together in the survey. Usually, this can be easily confirmed by subsequent
9
observations in consecutives waves. A very small number of individuals 12 disappear from the
survey for one or more interviews when still married or cohabitating and re-appear with a different
marital status (divorced or separated). For these couples, we assume they separate in the first year
that they are not observed in the survey 13. If a union ends, the partners are subsequently dropped
from the analysis sample. Also, couples in which the man is younger than 16 or older than 65 years
are dropped 14.
The analysis sample includes second and later marriages. Also we include both flow and stock
samples. The flow sample consists of marriages starting in 1991 or later while the stock sample
includes unions in existence at the start of the survey period. Models that distinguish between these
samples are estimated as part of the sensitivity analysis. Families formed before the beginning of
the survey can have idiosyncratically higher levels of durability and represent better matches. A
finding that job losses increase the probability of divorce even in families which are
idiosyncratically stable forms a conservative lower bound for the population at large.
Information on labour market behaviour and periods of unemployment is collected in different
sources within the BHPS. At each interview, the individual is asked about his/her current
employment situation 15, and whether he/she did any paid work or was away from a job in the week
prior to the interview. Retrospective information about labour force behaviour and all employment
spells over the previous year is also collected. G. Paull has compiled a special data set containing
labour forces spells (defined in terms of spell state, start date and end date) for each individual after
leaving fulltime education until the time of the interview (Halpin, 1997, Paull, 1997 and Paul
2002). This data set is complete for the first 11 waves of the BHPS and reconciles multiple sources
of information on employment spells.
The reason for the termination of an employment spell is not included in the Paull data set and was
derived from the job history files. When providing the reason for leaving a job, individuals can
choose among the following alternatives: promoted, left for better job, made redundant, dismissed
or sacked, temporary job ended, took retirement, stopped for health reasons, left to have a baby,
children/home care, care of other person, and other reasons. In this paper we focus on involuntary
displacements and consider only dismissals, redundancies and temporary job endings as job losses.
Also, only job losses experienced by the male partner are considered. There are a few reasons for
this. In many households men are the primary earners and their job loss will cause the largest
earnings shocks hence we are more likely to find impacts through that channel 16. Secondly, female
10
labour market mobility is much greater and due to a variety of reasons (e.g. child bearing and
rearing). Modelling these movements appropriately is complex. In the following we add a control
for the wife’s employment status hence our results are conditional on the employment profile of the
female partner. Our paper is a first step in a broader model of household behaviour that
incorporates both male and female job losses.
All involuntary job losses are expected to lead to negative shocks on earnings but dismissals are
more likely to incorporate individual traits and act as signals for the future match quality.
Temporary jobs are similar to dismissals in the sense that there may be an individual-specific
reason for the non-renewal of the contract; but it is also possible that the end date of the job was
fixed in advance (with no chance of renewal) in which case there is no signalling effect contained
in the termination of the employment spell (although there may be in the acceptance of such jobs).
The British redundancy law allows three reasons for redundancy: total cessation of the employer's
business (whether permanently or temporarily), cessation of business at the employee’s workplace
and reduction in the number of workers required to do a particular job. Moreover, the employment
law clearly specifies that, in a redundancy situation, the employer should select workers fairly and
should consider any alternatives to redundancy (this includes offering alternative work). Workers
are eligible for redundancy payments after two years of tenure on the job.
Despite its legal definition, redundancy can be used more generally as a term for involuntary
separation and respondents may be willing to report redundancies in cases of dismissals 17. This will
blur the distinction between dismissals and redundancies and inflate the effect of redundancies on
marital dissolution. We follow Borland et al. (1999) and treat redundancies differently based on the
industry of the job just ended. Specifically, data on industry-specific workforce growth rates are
collected from published UK government statistics and a variable measuring a three year moving
average growth rate for each industry is constructed. In some estimation models, job displacements
are separated depending on whether they refer to jobs in industries with declining or growing
employment.
Appendix Table 1 lists the explanatory variables (other than job displacements) used in the
empirical model. The choice of regressors follows the literature and includes human capital
indicators, income, children, and similarities between partners. These variables measure variations
in the utility of staying in the marriage, the value of the outside option, bargaining powers, and the
quality of the match. Income is measured as household non labour income and includes pensions
and other benefits, government transfers and investment income. The use of yearly income helps
11
smooth out effects of unusually high income receipt in any one month. Empirically, both yearly and
monthly incomes produce very similar results. Nonlabour income is included separately for the
wife and husband 18.
Other variables included are: age of husband, age of wife, highest educational qualification attained
(Degree, HND/A level, CSE/O level, No qualification), number of children in the household, a
binary indicator of the wife’s employment status and two match quality characteristics. The
economic literature related to marriage and divorce underlines the importance of “good matches”.
Couples are characterised by their “match quality” at the start of the relationship and this is an
important predictor of the future stability of their union. We include information about the
difference in age of the partners (a dummy variable equal to one if the difference in age is greater
than 8 years) and similarities in educational attainment (a dummy variable equal to one if the
partners have the same education category) to capture variations in match quality across couples.
For most models presented below the sample consists of 33463 observations involving 6134
couples. In terms of potential divorces, this sample covers the period 1993 to 2005. The reason for
the omission of 1991 and 1992 is as follows. All regressors are measured with a lag to prevent
reverse causality and because we do not know the exact timing of the divorce during the year. For
example, for someone who experienced a job loss between interviews in 1998 and 1999, we will
associate the occurrence of divorce or the continuation of the union between the interviews in 1999
and 2000. Hence, in the main sample, divorces can occur between 1993 and 2005 while the
exogenous variables are measured over the period 1992 to 2004. Furthermore, in several models we
include a second lag in the observed displacements (this is discussed below) so in these cases, the
estimation data include occurrences of job losses over the years 1991 to 2004.
Figure 1 displays the percentage rate of divorce/separation for couples who are in the analysis
sample. The sample is divided into 2 groups: those couples who experience at least one job loss
over the sample period and those who do not. From these raw numbers, we can see that on average
between 1 and 2% of unions are dissolved by divorce or separation each year and the incidence of
dissolutions trends slightly downwards over the length of the union. (Note that the average duration
of unions increases over time even in the unbalanced sample.) In total, 512 dissolutions are
observed in the sample.
Insert figure 1 here
12
Table 1 presents the number of job losses by year in the analysis sample. In total, there are 1,413
displacements consisting of 900 redundancies, 131 dismissals and 382 temporary job endings. If a
husband experiences more than one type of job loss in any year, this information is used in the
analysis 19. Generally, the incidence of displacements decreases over the 14 waves as the average
age of the sample rises. Exceptions occur around the recession of 2000-01 especially for
redundancies. In any one year, the incidence of job displacement is around 4 to 5%. This shows the
importance of large samples when studying the topic.
Figure 2 presents the distribution of length of marriages/cohabitations in the sample, by job loss
experience. From the raw figures we see that marriages with no job loss experiences are shorter on
average. This result is consistent with Charles and Stephens (2004) and would seem to contradict
the theoretical predictions discussed above. However, these figures do not take into account other
characteristics. Shorter marriages may have failed because of relatively bad match quality and since
these observations do not remain in the sample of couples, they are less likely to appear in the
sample that has experienced displacements. This illustrates the importance of controlling for
characteristics of the union and in particular, state dependence in the effect of marital duration.
Insert Figure 2 here
Table 2 presents sample means of demographic and socio-economic variables among couples with
and without job loss experience. Focussing on differences that are significantly different from zero,
we find that the sample of couples where husbands do not experience any job loss over the sample
period is slightly older (both partners), better educated (both partners) and the household nonlabour
income is higher. This is consistent with a stereotypical view of those households where partners
are relatively successful in the labour market and hold more secure jobs. Having the same
education level is more common in the sample of couples without any job loss, an indication of
better match quality. The table includes the divorce rate by sub-groups for the two samples. In
general, couples with a job loss experience also have higher divorce rates but most differences
between divorce rates by sub-groups are not significant. The highest divorce rates are found for
partners who are young and who have middle to low educational qualifications.
4. Estimation methods
We estimate a discrete time proportional hazards model, to investigate the effect of job loss on the
probability of a marital dissolution during time interval (t, t+1), given that the partnership has
13
survived until time t. A discrete time representation of the continuous time proportional hazards
model is given by:
ℎ𝑖 (𝑡) = Pr[𝑡 < 𝑇𝑖 ≤ 𝑡 + 1| 𝑇𝑖 ≥ 𝑡, 𝛽 ′ 𝑋𝑖 (𝑡), 𝛾(𝑡), 𝛼𝑖 ]
= 1 − exp[−exp{𝛽′ 𝑋𝑖 (𝑡) + 𝛾(𝑡) + 𝛼𝑖 }]
(1)
where t denotes time in the union, hi(t) is the hazard at time t for couple i (the dependence on X
and estimation parameters is suppressed), Xi(t) is a vector of covariates that potentially vary across
unions and time, β is a vector of coefficients common across time and unions, Ti is a discrete
random variable representing the time at which the union ends, γ(t) is the log of the integral of the
underlying continuous time baseline hazard between t and t+1. Variables and parameters are
assumed constant between t and t+1 for all t. We have made the usual assumption of common
effects of the covariates across unions; unobserved individual heterogeneity in the hazard model
takes the form of intercept shifts through the vector α which is further specified below. Finally, the
complementary log-log form of the hazard is implied by the underlying continuous time
proportional hazards specification.
To take into account censoring, the sample log-likelihood function of the observed duration data is
written with the aid of a dummy variable cit equal to 1 if 𝑡 < 𝑇𝑖 ≤ 𝑡 + 1 and the marriage is non-
censored (a divorce is observed during the time interval (t, t+1)) and cit = 0 otherwise (the marriage
continues on to the next interval or is censored). The log-likelihood function can be written as:
𝑇
ln 𝐿 = ∑𝑁
𝑖=1 ∑𝑡=𝜏𝑖 [𝑐𝑖𝑡 ln(ℎ𝑖 (𝑡)) + (1 − 𝑐𝑖𝑡 ) ln(1 − ℎ𝑖 (𝑡))]
(2)
where N denotes the number of couples in the sample, T is the maximum marital duration observed
in the sample, and τi equals 1 for the flow sample and the duration of the ith union in 1992 for the
stock sample (constructed using the starting date of the union). Independence between 𝑇𝑖 and cit
conditional on Xi(t) is a maintained assumption 20 (see Wooldridge, 2002, page 708).
The main issues in specifying this model revolve around the form of the baseline hazard and the
presence and form of couple-specific, time-invariant and unobserved effects denoted as α in (1).
These components of the model represent the systematic evolution of relationships over time (over
and above that captured by the covariates) and the presence of unobserved match quality
respectively. We adopt a flexible baseline that takes the form of a set of dummies γ(𝑡̃) equal to 1 if
the observation is in time interval 𝑡̃ and 0 otherwise. Ideally a full set of dummy variables (one for
14
each year covered in the sample) would be used. However due to small number of observations, we
group dummies over the following year intervals: (0 − 1, 2 − 4, 5 − 6, 7 − 10, 11 − 15, 16 −
20, 21 − 25, 26 − 30, 30+). As seen below the individual dummies are mostly insignificant with
this many categories (they are usually jointly significant), but since we do not care about the
individual coefficients we do not group them into coarser durations. An overly restrictive baseline
hazard will not fully take into account correlations in unobservables across time and lead to
inefficient estimates. As discussed next, we check for this with alternative approaches.
In the simplest models presented below, there are no couple-specific time-invariant unobserved
effects. It is well known that ignoring unobserved match-specific and time-invariant heterogeneity
will cause the overestimation of negative duration dependence since it then becomes the only form
of correlation over time in the model (other than that present in the covariates). However, since this
study focuses on the effects of job displacements on the duration of marriage, the distinction
between duration dependence (the shape of the baseline hazard) and unobserved heterogeneity is
secondary. Separating out unobserved heterogeneity from duration dependence does not change the
effects of the covariates on the mean duration (see Wooldridge, 2002, page 706). In other words,
the specification of a flexible baseline hazard capturing both duration dependence and unobserved
heterogeneity works just as well for our purposes. Nevertheless, models with flexible couplespecific time-invariant effects are useful as a specification check and we use the robustness of our
results as an indication of how well the baseline hazard is specified.
In the models with time-invariant unobserved heterogeneity, random effects independent of the
covariates and the censoring times are assumed 21. In the most flexible specifications, we use a
semi-parametric distribution based on the work of Heckman and Singer (1984) and Meyer (1990) 22.
Specifically the unobserved heterogeneity is assumed to be multiplicative in the hazard rate
specification:
ℎ𝑖 (𝑡) = 1 − exp[−exp{𝑥𝑖 (𝑡)′ 𝛽(𝑡) + 𝛾(𝑡) + ln(𝜀𝑖 )}]
(3)
where εi has a discrete distribution with a small number of mass points. The models with
unobserved heterogeneity shown below are estimated under the assumption of 2 mass points (a
model with 3 mass points is also estimated as a robustness check). An alternative model with
normally distributed random effects is also estimated and the results are qualitatively and
quantitatively very similar to those using the discrete distribution. More details are provided below.
15
Generally, flow samples are drawn from a population of short spells since their duration is limited
by the survey period (in our case 1991-2005). On the other hand, long spells are overly represented
in stock samples since spells that began and ended before the first interview are excluded. This is
the problem of length-biased sampling. In our analysis sample, these two subsamples are fairly
evenly divided: the stock sample contains 18279 observations (with 222 divorces and 697 job
losses) while the flow sample numbers 15184 (with 290 dissolutions and 716 job losses). In order
to check that the specification of the hazard (in particular the baseline hazard) is sufficiently
general to represent the non-random selection between stock and flow samples, we estimate models
that distinguish between the subsamples. Results are summarized below.
It is helpful to relate the econometric framework to a stylized marriage model where individuals
decide to stay married as long as the value of the match surpasses the outside option. Let q0i denote
the match quality at the time of marriage for couple i; q0i captures the partners’ knowledge of their
and their partner’s personality traits as well as their expectations regarding the future. The match
quality will evolve over time depending on choices (e.g. the decision to have children) and shocks
(e.g. involuntary job losses). Let qti represent the match quality at time t, then we can write qti =
f(q0i, q1i, ..., qt-1i, eti) where eti represents the innovation to match quality at time t. Similarly, the
outside option to the partnership, say zti, evolves over time depending on observable and
unobservable factors. A divorce will occur when qti - zti < 0. The probability of divorce at time t for
existing marriages -the hazard rate- is a function of match quality relative to outside options and
this in turn is dependent on the initial values and the innovations over time.
In the econometric framework explained above, the baseline hazard captures the evolution of the
match quality over time (relative to outside options) that is systematic across couples. The timeinvariant unobserved random effects capture the distribution of initial match quality (relative to
outside options) across couples and the explanatory variables measure the effects of couple and
time specific observables on the match quality relative to outside options.
As mentioned above our specification of the hazard links the probability of dissolution during a
time period t with control variables measured at t-1. We also include additional lagged observations
of job losses. This can be motivated as follows. Consider a person-specific dismissal. This event
implies a negative shock in earnings and a reduction in the value of the marriage. It also may signal
a shift in the perceived characteristics of the partner and a further reduction in the value of the
marriage. If this effect is only felt for a short time then including the one-period lagged dismissal is
sufficient. If the job loss implies a permanent drop in earnings or a permanent revision in the value
16
of the match then the job loss will permanently alter the probability of divorce. This means that all
past job losses should also be included in the specification of the divorce probability. Effects of job
losses that fade relatively quickly would require the inclusion of a few lags only. We experiment
with various specifications of the lagged job loss variables. Note that if, as expected, redundancies
do not convey signals regarding the partner’s traits, the impacts from longer lags in these
displacements should be small relative to other forms of displacements as they represent effects
caused by earnings losses only.
Taking into account the relative timing of the events does not necessarily control for all sources of
endogeneity of the job losses with respect to the probability of divorce. The issue arises because of
imperfect measurement of the match quality at the time of the job loss. In the context of our
duration model, any variation across couples in unobserved traits that are time invariant or that
evolve systematically over the length of the marriage will be controlled for by the couple-specific
effects and the baseline hazard. There remains the possibility of an unobserved, match-specific and
time-varying trait which is correlated with the occurrence of a job displacement. Consider a
particular worker who is chosen for dismissal because he has a particular character trait. If this
character trait is correlated with an unobserved component of the match quality (e.g. the trait
matters more because the match quality is poor relative to other similar couples and relative to
previous values of the match quality for this couple) then we will incorrectly attribute some of the
effect of the match quality to the job loss.
The possibility that a job loss captures both an income effect and a signal of an unfavourable
character trait does not necessarily lead to a problem for the causal interpretation of the impact of
job loss. The problem arises only when this character trait is correlated with omitted factors such as
unobserved match quality components. A job loss that doesn’t depend on the worker’s traits will
not pose this problem and our evidence suggests that this is the case for redundancies. As a
robustness check, we also estimate models excluding job losses other than redundancies in order to
see if the impact of redundancies remains stable after the omission of the possibly endogenous
dismissals and temporary job endings.
5. Results
In the estimation models presented below, the explanatory variables are measured at t-1 for divorce
risks at time t except where explicitly indicated. Estimates are presented in the relative risk format
(exp(β)); for an element of the coefficient vector β, say βk, these represent proportional changes in
17
the hazard due to the change of Xk from 0 to 1 in the case of a discrete variable or the addition of 1
unit to Xk if Xk is continuous. Standard errors are appropriately transformed. Generally, we present
standard errors that are robust and clustered by couple so that inferences are valid in the context of
a quasi-maximum likelihood framework where the standard distributional assumptions do not hold.
In the models with random effects, we present the usual standard errors. (In practise the robust and
clustered standard errors are very similar to the unadjusted ones, an indication that our specification
is sufficiently flexible in capturing correlations in durations across time.)
In Table 3 we present results on the job loss variables for a variety of specifications. (Results
concerning other regressors are presented below.) In the left-most column of results, the model
includes one job loss variable representing the incidence of (at least) one involuntary displacement:
either a dismissal, a redundancy or a temporary job ending. The effect is large; the probability of
divorce increases by 87% the following year. The average probability of divorce in any one year is
around 1 to 2% hence a job loss would raise this to about 3% on average for the sample. There is
evidence that the effect persists; with two lags, the probability of divorce is increased by 66% the
following year and 52% two years later. These are individually and jointly significant at the 1%
level.
The middle column shows the effects on dissolution of job losses separately for the different types
of displacements. Losing one’s job due to a dismissal has the highest effect on divorce: the
probability is increased by 154% one year later in the model with one lag. A job loss due to
redundancy has the smallest effect (39%) and it is not statistically significant. The effect of
temporary job endings is in between the two, and statistically significant. In the model with two
lags the coefficients on the redundancy variables are individually and jointly statistically
insignificant. Although positive, they are also quantitatively much smaller than those on other job
loss variables. For dismissals, only the effect of the first lag is statistically significant and it is
comparable to that found in the model that includes one lag only. Both dismissals and redundancies
have effects that taper off with time but this is not the case with temporary job endings. For this
reason we estimate a model with three lags in the job loss variables 23. In this case the coefficients
on temporary job endings fall as the lag increases and the coefficient on the third lag is individually
insignificant with a p-value of 0.841 24. (Results on the model with three lags are available upon
request.) Long term dynamic effects of job loss cannot be detected in these models 25.
The right-most columns correspond to a model where job loss variables are interacted with industry
group dummies. Industries are grouped based on whether a three year moving average workforce
18
growth rate is negative or positive and displacements from jobs in industries with declining
employment are treated separately from those located in growing industries. We expect that a
displacement in a declining industry will not include the same signalling content as an involuntary
job loss in a growing industry simply based on the probability of the event occurring.
The results for redundancies are again markedly different than for other job losses. The impact of
redundancies is larger and is significant in the declining industries. For the other job losses the
effects are also substantially larger and significant but only for the expanding industries (with the
exception of the second lag in the temporary job ending variable that is significant in both industry
groups). The probability of divorce is doubled for a couple in which the husband experienced a
redundancy from a declining industry while the probability is increased by over 160% when the
husband experienced a dismissal from an expanding industry.
In Table 4 we present results for semi-parametric models with random effects distributed discretely
with two mass points 26. (Models with normally distributed random effects yield very similar results
with coefficients generally differing at the second decimal point. These are available from the
authors.) The effects of this specification change on the estimated coefficients are marginal.
Without the industry group interactions, the only important change is that the dismissal dummy
lagged twice becomes marginally significant. With industry interactions, the temporary job endings
variable lagged twice becomes insignificant in declining industries. None of the general
conclusions regarding the job loss variables are altered. We can interpret this as an indication that
our model of the baseline hazard is flexible enough to capture most of the correlation across time in
unobservables.
Overall, the results in Tables 3 and 4 are consistent with the interpretations discussed above; that is,
redundancies represent mostly negative earnings shocks and hence the effects are smaller and more
short-lived. Dismissals are expected to have a larger signalling content regarding the future match
quality and hence a greater impact on the probability of dissolutions. Also since they capture
mostly the effects of negative earnings shocks, redundancies have worse effects in declining
industries compared to expanding industries given the difficulty of finding new and equivalent
employment in these sectors. The fact that dismissals and temporary job endings have more
substantial effects in expanding industries is also consistent with a signalling role for these
displacements; that is, unfavourable traits are more likely explanations of displacements in tight
labour markets.
19
Results with respect to other regressors for selected specifications are presented in Table 5. We
note that these estimates are similar across the columns and are generally insensitive to the specific
form used for the job loss variables. There are greater changes when random effects are added to
the model but qualitative results are unaffected. We find positive and significant effects from the
non labour income of the wife. The wife’s age shifts the hazard down while the dummy capturing a
difference in age between partners greater than 8 shifts the hazard up. Increasing both partners’ age
by one year shifts the hazard down (the exponential of the sum of the coefficients is 0.923) an
indication of greater marital stability with the age of the partners (note that this is over and above
the effects of the duration of the marriage). Shifts in education are insignificant in these models as
well as similarity of education across partners 27. It is interesting that in the British data we find no
significant effects of the number of children on the probability of divorce. In contrast, Charles and
Stephens (2004) found a negative effect of children on the probability of divorce in US data. The
baseline hazard dummies are generally individually insignificant although jointly they are
significant at a 5% level in all specification. (We discuss the shape of the baseline hazard further
below.)
Given our interpretation of dismissals and temporary job endings as capturing signals of
unfavourable traits, the question arises as to the possible endogeneity of these job loss variables. As
discussed above, if the trait being signalled is correlated with the match quality and if match quality
is imperfectly controlled for, then an endogeneity bias will result. Furthermore, we would expect
the bias to be positive in the sense that the strength of a signal of an unfavourable trait is likely to
be negatively correlated with match quality and hence increase the likelihood of divorce. In other
words the bias would increase the estimated impact of the job loss variables. Since the inclusion of
endogenous variables potentially biases all coefficients, we also estimate models where both
dismissals and temporary job endings have been omitted. The effects of redundancies are slightly
increased quantitatively, an indication of a slight correlation across types of job losses, but none of
our overall conclusions regarding the effects of redundancies are affected. (Results are available
upon request.)
Figure 3 plots the empirical hazard by year of marriage along with a predicted hazard. The
prediction is derived from the model with two lags and interactions with industry types although
the graph looks very similar across the various specifications. We see that the fit of the model is
quite good throughout the whole distribution of marriage length. Without controlling for sample
characteristics, the hazard is seen to decline over marital duration with a slight hump occurring
20
around 7 to 19 years of marriage. (For durations over 30 years, the numbers are too small to detect
any trend).
Insert figure 3 here
Figure 4 plots the baseline hazard for the same specification. The hazard is normalised to zero for
the reference duration segment (2 to 4 years). (Note that these values correspond to the unadjusted
coefficients while the results in the tables are presented in the relative risk format.) The risk of
divorce decreases over the first 7 to 10 years of marriage. This is followed by a sharp increase and a
flattening out at around 20 years. The risk of divorce declines after 25 years of marriage.
Individually, only the coefficient corresponding to the 16 to 20 years duration is significantly
different from zero although compared to the lowest risk at 7 to 10 years, the coefficients for the
durations 11 to 15, 16 to 20, and 21 to 25 are all individually significant. So is the coefficient for
the shortest durations of 0 to 1 years of marriage. The shape of the baseline suggests that
controlling for sample characteristics, the hump in the probability of divorce is sharper and occurs
later in the duration of the marriage than what is suggested by the unadjusted hazards.
Insert figure 4 here
Additional sensitivity analysis is described briefly in what follows. Detailed results are not
presented to save on space but are available upon request. We estimate models that include
interactions of displacements with recessionary years (2000-2001) on the grounds that these job
losses are less likely to convey any information on the partner’s traits. In general, the results
support our overall conclusions; however, we have small numbers of events (job losses during
recessionary years) and standard errors are large. Similar conclusions are reached with models that
include job losses after a marital dissolution as instruments.
Models in which coefficients are allowed to shift depending on whether the couple is part of the
stock or flow sample are estimated. As explained previously, we expect idiosyncratically long
(short) matches to form part of the stock (flow) sample. Specifically, variables are interacted with a
flow sample dummy. 28 The results involving the job loss variables described previously still hold
for the stock sample (the reference group) and with one exception, the interactions between the job
loss variables and the flow dummy are individually and jointly insignificant. The exception to this
is the interaction with the twice lagged temporary job ending that is marginally significant at 10%.
In contrast, the interactions involving variables other than the job loss variables are strongly jointly
21
significant. Also, the baseline hazard becomes flatter in these specifications as the flow dummy
now captures some of the duration effects. But as far as displacements are concerned, the overall
results presented above hold for both stock and flow subsamples.
The final robustness analysis incorporates self-reported satisfaction with one’s partner as a direct
measure of the match quality. Job displacements that are exogenous will not depend on the match
quality of the union. Regressions with job loss as the dependent variable and lagged satisfaction
variables as regressors are estimated. Unfortunately the satisfaction information is only available
starting in wave F (1996) of the BHPS and the sample size is substantially reduced. Nevertheless,
our results support the treatment of redundancies as exogenous in the sense that the satisfaction
measures do not help explain the incidence of job loss in the next period in the case of redundancies
in any of the models estimated. In contrast, the effects on temporary job endings and dismissals are
significant in some specifications. (Results are not shown to save on space but are available upon
request.) We are unaware of previous economic studies making use of satisfaction with one’s
partner in the context of the analysis of divorce risk and we believe this type of information to be a
promising avenue of future research.
6. Conclusion
In this paper we have examined the effects of involuntary job loss on partnership dissolution, a
topic that is particularly relevant in the current economic climate. Data from the British Household
Panel Survey are used in the study. We distinguish between different types of displacements
(dismissal, redundancy and temporary job ending) and we analyze their impacts on the probability
of divorce in the year following the job displacement. In general, job losses raise the probability of
divorce and these effects are stronger for dismissals and temporary job endings.
The evidence presented in the various specifications support the hypothesis that job losses that are
likely to depend on the worker’s characteristics contain signals of future match quality and hence
have a more important impact on the probability of match dissolution. Redundancies are dependent
on the employer’s characteristics and represent mainly earnings or psychological shocks. Their
impacts are smaller, shorter-term and have more influence in bad economic situations when the
earnings shock is expected to be more serious. The effects of redundancies are statistically
significant in a few of the models (for example if they occur in declining industries) but are usually
insignificant. In this sense, our results support those of Charles and Stephens (2004).
22
This analysis could be expanded in several directions. The analysis of the impacts of the wife’s job
loss is an obvious and interesting extension. Also, the role of social supports could be incorporated
by distinguishing the impact of job loss in high unemployment areas. Finally, the role of
expectations on job changes can be investigated using information on the worker’s opinion
regarding his/her job security29. Individuals who expect to lose their jobs may be paid
compensating wage differentials and these may partially protect the families from high distress and
other negative outcomes. On the other hand, they may make the negative earnings’ shock more
severe. The information on expectations currently available in the BHPS does not allow the
separation of voluntary job changes and hence could not be incorporated in this paper.
The finding of significant effects of job losses on the probability of divorce has important
consequences for the econometric modelling of the impacts of displacements on families generally.
Studies of the effect of job loss on family consumption or labour supply that only consider couples
who remain married will produce biased results since the couples who remain together are those
who had to face the fewest adjustments as a consequence of the loss of employment. Excluding
divorced couples is likely to lead to an underestimate of the impact of job displacements.
References
Arulampalam W. (2001) Is Unemployment Really Scarring? Effects of Unemployment
Experiences on Wages. The Economic Journal 111 (475): 585-606
Becker GS. (1974) A Theory of Marriage: part I. Journal of Political Economy 81 (4): 813-846
Becker, GS. (1974) A Theory of Marriage: part II. Journal of Political Economy 82 (2): S11-S26
Becker, GS., Landes E., Michael, R. (1977) An Economic Analysis of Marital Instability. Journal
of Political Economy 85 (6): 1141-1187
Becker, GS. (1991) A Treatise on the Family, Enlarged edition, Harvard University Press,
Cambridge MA
Blekesaune M. (2008) Unemployment and partnership dissolution. ISER working paper 2008-21.
Boheim R., Ermish J. (2001) Partnership Dissolution in the UK – the Role of Economic
Circumstances. Oxford Bulletin of Economics and Statistics 63(2): 197-208
Borland, J., Gregg P., Knight G., Wadsworth J. (1999) They get knocked down. Do they get up
again? Displaced workers in Britain and Australia. In: Kuhn, P. (ed.) Losing Work, Moving on:
lnternational Perspectives on Worker Displacement.W. E.Upjohn Institute for Employment
Research, Kalamazoo 301-374
Catalano R., Dooley D., Wilson, G., Hough, R. (1993) Job Loss and Alcohol Abuse – A Test Using
Data from Epidemiologic Catchment Area Project. Journal of Health and Social Behaviour 34(2):
215-225
Chan S., Stevens A. (2001) Job Loss and Employment Patterns of Older Workers. Journal of Labor
Economics 19(2): 484-521
23
Charles K.K., Stephens, M. (2004) Job Displacement, Disability and Divorce. Journal of Labor
Economics 22(2): 489-522
Clark. A., Oswald. A.J. (1994) Unhappiness and Unemployment. Economic Journal 104(424): 648659
Cullen J.B., Berry, J., Gruber, J. (2000) Does Unemployment Insurance Crowd out Spousal Labour
Supply?. Journal of Labor Economics 18(3): 546-572
Darity W., Goldsmith A. (1996) Social Psychology, Unemployment, and Macroeconomics. Journal
of Economic Perspectives 10(1): 121-140
Doiron D. (1995) Layoffs as Signals: The Canadian Evidence. The Canadian Journal of Economics
28(4): 899-913
Eliason M. (2004) Lost Jobs, Broken Marriages ISER Working Papers N. 21
Ercolani M., Jenkins S. (2000) The Labour Force Participation of Women Married to Unemployed
Men: is there an Added Worker Effect? Unpublished paper, Institute for Labour Research,
University of Essex
Ermisch, J. (2003) An Economic Analysis of the Family, Princeton University Press, Princeton
Gibbons R., Katz L. (1991) Layoffs and Lemons. Journal of Labor Economics, 9 (4): 351-380
Gould E., Paserman, D. (2003) Waiting for Mr. Right: Rising Inequality and Declining Marriage
Rates. Journal of Urban Economics 53 (2): 257-281
Halpin B. (1997) Unified BHPS Work-Life Histories: Combining Multiple Sources into a Userfriendly Format. Technical Papers of the ESRC Research Centre on Micro-Social Change,
Technical Paper 13
Heckman, J. and B. Singer (1984) A Method for Minimizing the Distributional Assumptions in
Econometric Models for Duration Data, Econometrica, 52(2): 271-320
Hill R. (1949) Families under Stress: Adjustment to the Crises of War Separation and Reunion,
Harpers, New York
Jacobson L., LaLonde R., Sullivan D. (1993) Earnings losses of displaced workers. American
Economic Review 83(3): 685-709
Jenkins S. (1995) Easy Estimation Method for Discrete Time Duration Model. Oxford Bulletin of
Economics and Statistics 57(1): 129-145
Jensen, P., Smith, N. (1990) Unemployment and Marital Dissolution. Journal of Population
Economics 3(3): 215-229
Kalil,
A.,
Ziol-Guest,
K.
(2007),
Parental
job
loss
and
academic progress in two parent families. Social Science Research 37(2): 500-515
children’s
Kyriacou D.N., Anglin D., Taliaferro E., Stone S., Tubb T., Linden J.A., Muelleman R., Barton E.
and Kraus J.F., (1999), Risk Factors for Injury to Women from Domestic Violence. New England
Journal of Medicine 341(25): 1892-1898
Kraft K. (2001) Unemployment and the Separation of Married Couples. Kyklos 54(1): 67-88
Lam D. (1988) Marriage Markets and Assortative Mating with Household Public Goods:
Theoretical Results and Empirical Implications. Journal of Human Resources 23 (4): 462-487
Lundberg S., Pollak R.A. (2006) The American Family and Family Economics. Journal of
Economic Perspectives 21(2): 3-26
Manser M., Brown, M. (1980), Marriage and Household Decision-Making: A Bargaining Analysis.
International Economic Review 21(1): 31-44
24
Matouschek N., Rasul I. (2008), The Economics of the Marriage Contract: Theory and Evidence.
Journal of Law and Economics 51(1): 59-110
McCubbin, H.I., Patterson, J. M. (1982) The Family Stress Process: the Double ABCX Model of
Adjustment and Adaptation. In McCubbin H.I., Cauble AE. , Patterson JM. (eds.), Family Stress,
Coping, and Social Support, Haworth Press, Springfield 169-188
McElroy M.B., Horney M.J. (1981) Nash-Bargained Household Decisions: Toward a
Generalization of the Theory of Demand. International Economic Review 22(2): 333-349
Meyer, B. D. (1990). Unemployment insurance and unemployment spells. Econometrica 58(4):
757-82
Miller, M.A., Rahe R.H. (1997) Life Changes Scaling for the 1990’s. Journal of Psychosomatic
Research 43(33): 279-92
Paull, G. (1997), Dynamic Labour Market Behaviour in the British Household Panel Survey: The
Effects of Recall Bias and Panel Attrition, The Labour Market Consequences of Technical and
Structural Change. Discussion Paper Series no. 10, Centre for Economic Performance
Paull, G. (2002), Biases in the Reporting of Labour Market Dynamics, The Institute of Fiscal
Studies. Working Paper Series, No. 02/10
Prietro-Rodriguez, J. and Rodriguez-Gutierrez, C. (2003), Participation of Married Women in the
European Labour Markets and the "Added Worker Effect”. Journal of Socio-Economics 32(4):
429-446
Pronzato C. (2007) Family histories from BHPS. British Household Panel Survey Consolidated
Marital, Cohabitational and Fertility Histories, 1991 – 2005. Downloadable from UK Data Archive
Ruhm C. (1991) Are Workers Permanently Scarred by Job Displacements? American Economic
Review 81(1): 319-323
Schwerdt, G. (2007) Labor Turnover Before Plant Closure: ‘Leaving the Sinking Ship’ vs ‘Captain
Throwing Ballast Overboard.’ EUI Working Paper ECO 2007/22
Stephens M. (2001) The Long Run Consumption Effects of Earnings Shocks. The Review of
Economics and Statistics 83(1): 28-36
Stephens M. (2004), Worker Displacement and Added Worker Effect. Journal of Labor
Economics, 20(4): 504-537
Strom S. (2003) Unemployment and Families: A Review of Research. Social Service Review 77(3):
399-430
Sullivan D., Watcher T. (2006, preliminary) Mortality, Mass Layoffs ad Career Outcomes: an
Analysis using Administrative Data. Working Paper Series WP-06-21, Federal Reserve Bank of
Chicago
Voydanoff P. (1990) Economic Distress and Family Relationships: a Review of the Eighties.
Journal of Marriage and the Family 52 (4): 1099-1115
Weiss, Y., Willis, R. (1997) Match Quality, New Information and Marital Dissolution. Journal of
Labor Economics 15(1): 629-79
Weiss Y. (2000) Economic Theory of Marriage and Divorce. In N. Smelser and B. Baltes (eds),
International Encyclopedia of Social and Behavioral Sciences, Amsterdam
Winkelman, L. Winkelmann, R. (1998) Why are Unemployed so Unhappy? Evidence from Paned
Data. Economica 65(257): 1-15
Wooldridge J. (2002) Econometric Analysis of Cross Section and Panel Data, The MIT Press,
Cambridge, Massachussets
25
Figure 1 – Divorce rate by year, analysis sample
3.50%
3.00%
2.50%
2.00%
Job loss sample
1.50%
No job loss sample
1.00%
0.50%
0.00%
The sample contains 33463 observations involving 6137 couples. Divorce includes separation.
26
Figure 2 – Distribution of years of marriage, analysis sample
16.00%
14.00%
12.00%
10.00%
8.00%
No Job loss sample
6.00%
Job loss sample
4.00%
2.00%
0.00%
Years of marriage
The sample contains 33463 observations involving 6137 couples. Marriage includes cohabitation.
27
Figure 3 – Observed and predicted hazards by duration of marriage
0.035
0.030
0.025
Observed
Predicted
0.020
0.015
0.010
0.005
0.000
Years of marriage
The predicted probability of divorce is generated from the model presented in the right-most columns
of Table 3.
28
Figure 4 – Baseline hazard
0.014
0.012
0.01
0.008
0.006
0.004
0.002
0
Years of Marriage
The predicted probability of divorce is generated from the model presented in the right-most columns
of Table 3.
29
Table 1- Involuntary job displacements by year and type
Year
No.
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
6
11
13
8
13
10
11
12
10
21
11
3
2
Total
131
Dismissals
% of
% of
total
couples
4.58
0.28
8.40
0.54
9.92
0.63
6.11
0.40
9.92
0.61
7.63
0.48
8.40
0.47
9.16
0.56
7.63
0.40
16.03
0.53
8.40
0.34
2.29
0.08
1.53
0.06
100.0
0.39
No.
88
99
80
57
61
54
43
56
67
166
52
45
32
900
Redundancies
% of
% of
total
couples
9.78
4.12
11.00
4.82
8.89
3.88
6.33
2.82
6.78
2.88
6.00
2.58
4.78
1.83
6.22
2.63
7.44
2.67
18.44
4.23
5.78
1.62
5.00
1.25
3.55
0.98
100.0
2.69
Temporary job endings
% of
% of
No.
total
couples
20
5.24
0.94
32
8.38
1.56
34
8.90
1.65
26
6.81
1.29
28
7.33
1.32
24
6.28
1.15
38
9.95
1.62
28
7.33
1.32
42
10.99
1.68
68
17.80
1.73
22
5.76
0.68
13
3.40
0.36
7
1.83
0.22
382
100.0
1.14
The sample size is 33463. See the main text for definitions of the types of displacements. % of couples is
the proportion of job losses relative to the number of couples in the analysis sample in the year in
question. For the total, the % of couples refers to the proportion of displacements in the total number of
observations.
30
Table 2 - Means of socio-economic variables and divorce rates by job loss samples. (33463 obs.)
Job loss sample
Sample mean
Divorce rate
No job loss sample
Sample mean
Divorce rate
Men - Education
High degree*
11.59†
1.78
15.13
1.24
HND/A level
43.21
1.78
42.21
1.48
CSE/O level
21.30
2.32
21.25
2.10
No qualification
23.90†
1.21
21.41
0.96
- Age
18-30
14.75
3.82
14.15
3.94
31-40
29.12
2.31†
29.17
1.58
41-55
39.27†
1.02
37.69
0.99
56-65
16.86
0.73
18.99
0.38
43.64
Mean age in years
43.20†
Women - Education
High degree*
9.41†
1.61
12.87
1.31
HND/A level
32.79†
1.80
36.05
1.64
CSE/O level
34.14†
2.21
30.68
1.69
No qualification
23.65†
1.10
20.40
0.92
- Age
18-30
20.09
3.56
19.28
3.51
31-40
29.96
2.06
30.42
1.52
41-55
38.85†
1.06
36.42
0.80
56-65
11.09†
0.12
13.90
0.27
41.06†
41.58
Mean age in years
- Work status
In paid employment or self employed
67.72
1.60
68.00
1.50
Unemployed, retired, family care, other*
32.28
2.09†
32.00
1.40
Household non labour income
0-1,000
34.35
1.36
34.80
1.75
1,001-5,000
41.02†
1.87
38.57
1.48
>5,000
24.63†
2.12†
26.63
1.09
3791.50†
4307.17
Mean income in £ (base year=2005)
0.90
0.88
Number of children
Couples with children
46.76
2.35†
45.61
1.70
Couples without children
53.24
1.24
54.29
1.27
Partners have same education levels
Yes
40.67†
1.86
44.23
1.37
No
59.33†
1.54
55.77
1.69
Difference in partners’ age >= 8 years
Yes
8.81
3.12
9.31
2.01
No
91.19
1.63
90.69
1.41
The number of observations is 7276 for the job loss sample and 26187 for the no job loss sample. All
figures are percentages unless otherwise indicated. * denotes an omitted group in regressions. †
denotes that the difference in the statistics from the job loss and no job loss samples is significantly
different from zero based on a two-tailed test and a 5% level of significance. See Appendix Table 1 and
the main text for more details on the variables. For age, income and number of children, only the
continuous variable is included in the estimation models.
31
Table 3 – Effects of job loss on the probability of dissolution. Sample size = 33463.
Any job loss
Job loss by type
Job loss by industry group
Declining
Variables
Expanding
Exp(coef) (st.err.) Exp(coef) (st.err.) Exp(coef) (st.err.) Exp(coef) (st.err.)
a) Models with one lag only
Any job loss t-1
Redundancy t-1
Temp job end t-1
Dismissal t-1
Log pseudolikelihood
No. of parameters
1.867*** (0.309)
-2468.40
24
1.389
(0.316)
2.126*** (0.524)
2.537*** (0.911)
-2465.98
26
2.177** (0.699) 1.070
(0.352)
1.626
(0.828) 2.433*** (0.693)
2.049
(1.499) 2.499** (1.112)
-2463.38
31
1.350
(0.303)
1.041
(0.242)
1.696* (0.427)
1.834* (0.444)
2.151* (0.885)
1.572
(0.612)
-2462.34
29
2.000**
0.777
1.191
2.071*
1.725
1.749
b) Models with two lags
Any job loss t-1
Any job loss t-2
Redundancy t-1
Redundancy t-2
Temp job end t-1
Temp job end t-2
Dismissal t-1
Dismissal t-2
Log pseudolikelihood
No. of parameters
Wald tests:
H0: all job loss coeffs=0
H0: t-1 job loss coeffs=0
H0: t-2 job loss coeffs=0
H0: redundancy coeffs=0
1.664*** (0.283)
1.522*** (0.242)
-2465.23
25
χ2(2)=20.48
p-value=0.000
χ2(6)=29.11
p-value=0.000
χ2(3)=9.54
p-value=0.023
χ2(3)=7.94
p-value=0.047
χ2(2)=1.82
p-value=0.403
(0.632) 1.093
(0.284) 1.311
(0.628) 1.998**
(0.859) 1.880**
(1.315) 2.614*
(1.063) 0.824
-2459.07
38
χ2(6)=13.91
p-value=0.031
χ2(3)=5.85
p-value=0.119
χ2(3)=5.83
p-value=0.120
χ2(2)=5.06
p-value=0.080
(0.363)
(0.438)
(0.587)
(0.588)
(1.330)
(0.549)
χ2(6)=18.71
p-value=0.005
χ2(3)=8.88
p-value=0.031
χ2(3)=4.51
p-value=0.212
χ2(2)=0.50
p-value=0.779
All regressions include the following variables: age of husband, age of wife, husband and wife’s education (3 dummies
for each partner), wife’s employment status (one dummy), husband’s nonlabour income, wife’s nonlabour income,
number of children present in household, dummy for difference in age of partners greater than 8, dummy for similar
education level and 8 dummies for duration of marriage. A constant is included. Models that distinguish between
industry groups also include missing industry dummies interacted with redundancies (2 dummies) and temporary job
endings (1 dummy); these are insignificant in all models. Regressions are cloglog models estimated in Stata. Standard
errors are robust and clustered by couple. * indicates that exp(coeff) is significantly different from 1 at 10% level, ** at
5% and *** at 1%.
32
Table 4 – Effects of job loss on the probability of dissolution - random effects models.
Sample size = 33463.
Job loss by type
Job loss by industry group
Declining
Variables
Redundancy t-1
Redundancy t-2
Temp job end t-1
Temp job end t-2
Dismissal t-1
Dismissal t-2
Mass pt 1: location, prob.
Mass pt 2: location, prob.
Log likelihood
No. of parameters
Exp(coef) (st.err.)
Exp(coef) (st.err.)
1.353
(0.325)
1.061
(0.256)
1.971** (0.567)
1.811** (0.486)
2.080* (0.847)
1.875* (0.711)
-1.653, 0.972
1.970, 0.027
-2422.16
31
1.931***
1.338
1.300
2.095
1.807
1.966
Expanding
Exp(coef) (st.err.)
(0.661)
1.127
(0.447)
0.772
(0.768)
2.408***
(0.974)
1.866*
(1.499)
2.489*
(1.302)
1.003
-1.671, 0.971
1.904, 0.029
-2419.93
40
(0.404)
(0.300)
(0.772)
(0.602)
(1.270)
(0.580)
All regressions include the following variables: age of husband, age of wife, husband and wife’s education (3 dummies
for each partner), wife’s employment status (one dummy), husband’s nonlabour income, wife’s nonlabour income,
number of children present in household, dummy for difference in age of partners greater than 8, dummy for similar
education level and 8 dummies for duration of marriage. A constant is not included. Models that distinguish between
industry groups also include missing industry dummies interacted with redundancies (2 dummies) and temporary job
endings (1 dummy); these are insignificant in all models. Models include random effects with discrete distributions
containing 2 mass points. All models are estimated with gllamm routines written for Stata. Standard errors are robust
and clustered by couple. * indicates that exp(coeff) is significantly different from 1 at 10% level, ** at 5% and *** at
1%.
33
Table 5 – Hazard models of the probability of dissolution- other regressors. Sample size = 33463.
Two lags, industry
Job loss has
Two lags and
groups and
One lag
two lags
industry groups
random effects
Variables
Age:
Husband
Wife
Education:
Husband – HND/A
Husband – CSE O
Husband – No qual
Wife – HND/A
Wife – CSE O
Wife – No qual
Nonlabour inc.:
Husband
Wife
Wife employed
Number of children
Match quality:
Age Diff > 8 yrs
Same educ level
Baseline hazard:
Year 0 to 1
Year 5 to 6
Year 7 to 10
Year 11 to 15
Year 16 to 20
Year 21 to 25
Year 26 to 30
Year 31 +
Log pseudolikelihood
No. of parameters
Wald test of H0:
baseline coeffs=0
Exp(coef) (st.err.) Exp(coef) (st.err.) Exp(coef) (st.err.) Exp(coef) (st.err.)
0.975
(0.016)
0.947*** (0.016)
0.975
(0.016)
0.947*** (0.016)
0.976
(0.016)
0.947*** (0.016)
0.976
(0.016)
0.940*** (0.015)
1.119
1.305
0.995
1.151
1.257
1.176
(0.192)
(0.239)
(0.205)
(0.204)
(0.240)
(0.273)
1.118
1.296
0.993
1.160
1.264
1.181
(0.192)
(0.238)
(0.204)
(0.206)
(0.242)
(0.274)
1.129
1.307
1.005
1.144
1.243
1.165
(0.194)
(0.239)
(0.207)
(0.203)
(0.238)
(0.271)
1.166
1.333
1.000
1.133
1.172
1.054
1.695
4.852***
0.895
0.920
(0.703)
(2.258)
(0.100)
(0.048)
1.688
4.818***
0.934
0.922
(0.707)
(2.250)
(0.110)
(0.048)
1.686
4.858***
0.914
0.924
(0.703)
(2.263)
(0.103)
(0.049)
1.388
(1.246)
6.087*** (4.060)
0.934
(0.110)
0.913
(0.051)
(0.211)
(0.260)
(0.226)
(0.215)
(0.235)
(0.255)
1.470** (0.283)
0.910
(0.086)
1.474** (0.283)
0.908
(0.086)
1.464** (0.281)
0.901
(0.086)
1.606** (0.352)
0.936
(0.101)
1.173
(0.179)
0.905
(0.145)
0.812
(0.128)
1.289
(0.203)
1.446* (0.278)
1.446
(0.333)
0.978
(0.333)
0.736
(0.295)
-2465.98
26
χ2(8)=16.18
p-value=0.040
1.153
(0.175)
0.908
(0.145)
0.816
(0.129)
1.289
(0.203)
1.442* (0.277)
1.447
(0.333)
0.973
(0.331)
0.732
(0.293)
-2462.34
29
χ2(8)= 15.76
p-value=0.046
1.172
(0.178)
0.913
(0.146)
0.816
(0.129)
1.287
(0.204)
1.435* (0.277)
1.446
(0.333)
0.970
(0.330)
0.731
(0.293)
-2459.07
38
χ2(8)=15.73
p-value=0.046
0.978
(0.170)
1.113
(0.187)
1.037
(0.180)
1.628*** (0.300)
1.791*** (0.399)
1.895
(0.494)
1.111
(0.387)
0.852
(0.349)
-2419.93
40
χ2(8)=15.90
p-value=0.044
All regressions include a constant (except for the model with random effects) and job loss variables by type of
displacement. Results on job loss variables are presented in Table 3; the models can be matched by the value of the
pseudolikelihood. Regressions without random effects are cloglog models estimated in Stata; unobserved
heterogeneity is modelled as random effects discretely distributed with two mass points and this model is estimated
with gllamm routines written for Stata. Standard errors are robust and clustered by couple except for the random
effects model. * indicates that exp(coeff) is significantly different from 1 at 10% level, ** at 5% and *** at 1%.
34
Appendix
Appendix Table 1 - Variable definition - additional regressors.
Husband’s age
Wife’s age
Difference in age
Education: Degree
HND/A
O/CSE
No qualification
Similarity in education
Husband’s non labour income
Wife’s non labour income
Number of children
Wife’s employment
Years
Years
1 if the difference in age across partners >= 8 years
1 if the highest academic qualification is a university degree. This is the
omitted category.
1 if the highest academic qualification is HND (including teaching
qualification, nursing or other higher qualification) or GCE A level
(upper high school graduate)
1 if the highest academic qualification is GCE O level or CSE (lower high
school graduate).
1 if highest qualification is less than high school
1 if partners have the same highest qualification
In ‘000 £ (deflated using 2005 as base year).
In ‘000 £ (deflated using 2005 as base year).
Number of dependent children in the household
1 if the wife is in paid employment or self employed
35
1
We thank participants of the 2007 Australian Labour Market Research workshop and the 2008 Australian Conference
of Economists for their suggestions and comments. Special thanks to two anonymous referees and associate editor
Deborah Cobb-Clark for valuable comments and suggestions. The BHPS data was provided by the Economic and
Social Research Council’s Data-Archive at the University of Essex and is used with permission. The usual disclaimer
applies.
2
Several studies of the effects of job displacements on earnings have used plant closures as exogenous displacements
(see for example Gibbons and Katz, 1991 for the US and Doiron, 1995 for Canada). There is some dispute about the
treatment of plant closures as exogenous (Schwerdt, 2007). In any case, information on plant closures is not available
in the BHPS.
3
The number of observations with displacements due to plant closures is not provided but the estimation results (some
substantial quantitative effects with large standard errors) suggest that perhaps the number of such job losses is too
small to yield sufficient precision in the estimates.
4
Eliason uses propensity score matching to compare two samples of married individuals with one sample consisting of
persons who have experienced a plant closure during the year 1987. Length of marriage is used as a matching variable
but this variable only distinguishes between unions of less than 3 years.
5
See also Eliason (2004) for a more detailed explanation of this model.
6
A related strand of the literature considers the impact of earnings’ shocks generally on family consumption and
production. See for example Browning and Crossley (2001) and Cullen and Gruber (2000).
7
See Darity and Goldsmith (1996) for a review.
8
Studies have also found negative impacts of unemployment on the well-being of spouses and children. Most of these
papers are also found in other fields of study such as psychology and social sciences (see Strom, 2003; Voydanoff,
1990 and Kalil and Ziol-Guest, 2007).
9
Additional samples of 1,500 households in each of Scotland and Wales were added in 1999, and in 2001 a sample of
2,000 households was added in Northern Ireland, making the panel suitable for UK-wide research. These samples are
included in our analysis.
10
In our main sample, 14% of observations consist of cohabitations. The modeling of the duration of the union
distinguishing by type of union is not straightforward unless one assumes independent shocks (or competing risks) and
although an interesting extension, it is left for future work.
11
We would ideally like to know the date at which individuals felt their marriage end, regardless of the legal date of
divorce or separation but this is not easily defined.
12
Less than 25 couples.
13
A sensitivity analysis is conducted by constructing a binary variable for couples who disappear from the survey and
re-appear with a different marital status. This variable is introduced in our main models and does not affect the sign and
significance of job loss variables. Results are available on request.
14
Those couples where the man reaches 65 during the survey period are dropped at the time the man reaches 65 and
treated as right-censored. We use age 65 as an exogenous censoring device. Due to the presence of mandatory
retirement, the role of job displacements for workers older than 65 is likely to be quite different than for younger
individuals.
15
The proposed alternatives are: self employed, in-paid employment (full time or part time), unemployed, retired from
paid work, on maternity leave, looking after family or home, full time student/at school, long term sick or disabled, on
a government training scheme, something else.
16
In our main sample, 72% of households report the husband as the individual with the largest labour income.
17
See Borland et al. (1999).
18
Previous research suggests that spousal labour income may be endogenous to job displacement so the wife’s labour
income is not part of the main model.
19
There is a limited incidence of repeated job loss of the same type in the same year mostly involving temporary job
endings. Specifically, out of all observations with either a dismissal or redundancy (1480 couple - year), 106 or 7%
have more than one occurrence of the job loss. Not surprisingly, the number is a lot higher for temporary job endings
(21%). Sensitivity analysis is conducted with the addition of dummies for the observations with multiple occurrences
and results are very similar to those reported below. Details are available from the authors.
20
This assumption rules out right censoring rules that are correlated with unobservables and makes the use of selfreported disability or early retirement as censoring variables problematic. In our sample, durations are right-censored
when they reach the end date of the sample, when the husband’s age reaches 65 or due to attrition from the sample.
21
With single spell data, it is very difficult to allow for correlations between the time-invariant unobserved
heterogeneity and the covariates.The independence assumption which must be maintained with random effects has
implications for the measurement of the impact of the job displacement variables. Specifically, in models with
unobserved time-invariant random effects, any signal contained in a job loss variable must be independent of any initial
unobserved match value. One can easily imagine violations of this assumption. A strong marriage may be harder to
influence; hence the signalling effect of a dismissal may be lower for these couples. Again we stress that these results
are used more as sensitivity analysis, in particular as a check on the specification of the baseline hazard.
36
22
We are grateful to an anonymous referee for suggesting this approach.
Adding a third lag reduces the sample to 27667 observations.
24
The third lags on the displacement variables are also jointly insignificant; a Wald test yields a χ2(3) value of 1.60
which corresponds to a p-value of 0.659.
25
Charles and Stephens (2004) also found that the effects of job displacements were short run. In their specification,
displacements in the previous 3 years were grouped and these had significant positive effects on the probability of
divorce. There was no evidence of effects for those job losses that occurred in the previous 4 to 5 years. For job losses
that occurred more than 5 years ago, a negative effect was found in the case of layoffs but no effects were detected for
displacements due to plant closures. They interpret the long term effects from layoffs as an indication that the
marriages involved survived a crisis and came out with a strengthened relationship. They also argue that the lack of
effects in the medium term following a displacement can be perceived as evidence that the effects they do find (either
from plant closures or layoffs) cannot be due to an omitted (time-invariant) variable. In the context of our paper, the
omission of a time-invariant effect also cannot explain the effects of job displacements since results from the random
effects model are virtually the same as those of the main model.
26
We estimate models with three mass points but these did not converge easily; specifically, we had to omit the age
variables and restrict the baseline hazard to get convergence. In all cases, the probability of the third mass point was
between 0.011 and 0.012 and the results on the job loss variables were similar to those presented in Table 4 except for
coefficients on the temporary job endings that became smaller and generally insignificant.
27
We should add that results on education, nonlabour income and the wife’s employment status are sensitive to the
treatment of the age variables. This is not surprising given the correlation in these variables. Since we do not care about
these variables per se we choose the more flexible specification and include all regressors.
28
Interactions with the baseline hazard were restricted to due to the short marriages in the flow sample.
29
This was suggested by an anonymous referee.
23
37